Saturday, November 21, 2026

Every AI Model Has to Make Value Choices

Arrow’s impossibility theorem suggests the impossibility of aligning human values with any single artificial intelligence model, for the simple reason that choices must be made about such values. 


In other words, whose values should AI agents be aligned with? Take your pick: females or males; urban or rural; old or young; liberal or conservative; religious or atheist; rich or poor; European, African, Asian, Latin American or Native American? 


Any model developer has to develop weights that make choices between such assumptions. The Arrow Impossibility Theorem suggests that universally aligning a single AI model to diverse human values is mathematically impossible, forcing AI developers to make structural trade-offs. 


One can see this in the “paradox of voting,” dealing with ranked choices, discovered by the Marquis de Condorcet in 1785, for example. 


That principle shows that possibilities for choosing rationally can be lost when individual preferences are aggregated into social preferences. 


Likewise, Arrow’s theorem suggests that collective decisions cannot be made on the basis of a common preference that assimilates the tastes and values of all the individual men and women who make up a society. Not hard, “impossible.” 


The question that Arrow’s theorem answers is: which procedures are there for arriving at a social ordering of some given alternatives, on the basis of some given people’s preferences among them, no matter what these preferences turn out to be?


And the answer turns out to be that there are no reliable procedures. Systems of ranked choice voting can produce outcomes that reflect “nobody’s” actual personal preferences, for example. Given three choices, A, B and C, it is possible that the ranked choice system produces an outcome nobody actually desired. 


Say there are three alternatives: A, B and C, and three choosers. Each ranks their own preferences::


ABC

BCA

CAB. 


In this scenario, there is no way to reconcile the differing “highest” value choices. One examines the second and third choices, looking for preference there. It cannot be found. So then one tries to assess the rankings any other way that attempts to assess which alternative “best” reflects a blended choice.


Looking at the sum total of first, second and third choices, no “clear” pattern emerges. 


Suppose, taking the alternatives pair by pair, we put the matter to a vote: we count one alternative as socially preferred to another if most voters prefer it. We determine in this way that A is socially preferred to 

B, since two voters (1 and 3) prefer A to B, but only one (voter 2) prefers B to A. 


Similarly, B is socially preferred to C. We might then expect that A is socially preferred to C. 


But there also are two voters who prefer C to A. So one winds up with a choice nobody actually supports. 


In the context of model training, similar choices have to be made, when there are competing and conflicting views. 


Since people, in addition to having different values and opinions, also have different intensities of feeling about their many values, there is no way of accounting for how much more strongly one must prefer one thing to another in order to make up for the fact that someone else’s preference is just the other way around. 


So the problem is that there is no way to satisfy all of four conditions:

  • Universal domain: the aggregation rule should work for every possible profile of individual preferences. We cannot exclude unusual preferences in advance.

  • Pareto efficiency: if every individual strictly prefers a to b, the social ordering should too. This is the bare minimum of respecting individual preferences.

  • Independence of irrelevant alternatives: the social ranking of a versus b should depend only on individual preferences over a versus b, not on preferences over some third alternative c.

  • Non-dictatorship: there should be no single individual whose preferences always determine the social ordering.


The upshot is that models cannot be trained in a way that represents the full and whole group. So model developers use workarounds that are compromises. 


Workaround

How It Bypasses Arrow's Theorem

AI Implementation

Narrow / Pluralistic Alignment

Abandons the goal of one universal ranking for everyone.

Steerable models, custom system prompts, or community-specific models tailored to distinct user groups.

Game-Theoretic Alignment

Replaces a single scalar reward score with equilibrium strategies.

Nash Learning from Human Feedback (NLHF), where the model learns a mixed strategy that cannot be beaten by any alternative response.

Cardinal / Fine-Grained Feedback

Uses detailed ratings or multi-objective scores instead of simple pairwise rankings.

Training separate reward models for distinct axes (helpfulness, harmlessness, honesty) and letting users or routers weight them.


Some of us might argue that only Grok is explicit about its own choices. The others might claim they are neutral or objective, when in fact they must make some value choices when training their models, just as 

Grok does.


Thursday, November 19, 2026

AI Investment Propels U.S. GDP Growth; Productivity Impact Starts to Emerge

First-half U.S. gross domestic product growth without computing-related spending would have been just 0.1 percent, calculates Harvard economist Jason Furman. U.S. GDP growth in the first half of 2025 was almost entirely driven by investment in data centers and information processing technology, he estimates.

source: Jason Furman 


In 2026, full year impact ranges from 14 percent to 44 percent.  


To be sure, gross investment does not account for the incidence of such investments, as many components and products are imported, and may not directly increase U.S. GDP. A large share of equipment (servers, chips, for example) is imported, so it appears in investment (positive) and imports (negative), largely canceling for GDP. 


Structures (data center construction), software, and research and development are more domestically produced and contribute more fully to GDP metrics. 


source: Federal Reserve


Comparing expected AI infrastructure investments to past waves of infra building, AI might surpass many others as a percentage of gross domestic product, notably becoming three times bigger than the turn of the century investment in optical fiber facilities, for example. ==============

source: Wall Street Journal 


And there is now some evidence of productivity growth related to use of AI. 



Study and date

Methodology / population

Findings

Significance

MIT  / Microsoft / Accenture / Fortune 100, February 2026

Three randomized field experiments involving 4,867 software developers

AI coding assistants increased completed tasks by 26.08% overall. Less experienced developers experienced larger gains.

Strong causal evidence of worker-level productivity improvements.

BEA, Early Estimates of the Impact of AI, February 2026

Econometric analysis of BEA industry economic accounts

Baseline model finds AI associated with higher productivity and reduced input use, including labor. Results are sensitive to assumptions about AI adoption timing.

Particularly important because it attempts to identify AI effects in national economic statistics.

Kansas City Federal Reserve , A New U.S. Productivity Chapter?, February 2026

Industry productivity data compared with Census AI-adoption data

U.S. labor productivity has risen above its pre-pandemic trend. Higher AI adoption is associated with faster industry productivity growth, but AI explains little of the aggregate improvement.

Important evidence both for and against an emerging AI productivity boom.

NBER, Artificial Intelligence, Productivity, and the Workforce, March 2026

Survey of nearly 750 corporate executives

More than half of firms had invested in AI. Productivity gains were positive but heterogeneous, with larger effects in high-skill services and finance.

Suggests AI is beginning to affect business-level productivity, but with delayed revenue realization.

U.S. Census Bureau , The Microstructure of AI Diffusion, April 2026

Nationally representative Business Trends and Outlook Survey, supplemented with firm and worker-task analysis

18% of firms used AI in a business function; 32% when weighted by employment. Broader AI integration was positively associated with commercial performance.

Useful evidence about the actual breadth and depth of adoption.

Generative AI and Labour Productivity: A Quasi Experiment on Coding, June 2026

Quasi-experiment involving software developers using Ant Group's CodeFuse

CodeFuse adoption increased code output by more than 50%.

Strong additional evidence of substantial productivity gains in software development, although outside the U.S.

Federal Reserve, The Potential for Sustained Productivity Impetus from GenAI, September 2026

Economic analysis of general-purpose technologies and innovation processes

Finds suggestive evidence that GenAI could function both as a general-purpose technology and as an invention-method improvement, increasing future productivity growth.

Important argument for potentially sustained, rather than temporary, productivity improvements.


Still, some skeptics will continue to note that gains are seen most clearly at the individual worker and firm level, not in the statistics on gross domestic product. Some of us would argue that is entirely keeping with the history of productivity gains from deploying important new technologies. 



So one wonders whether policymakers will be so set on slowing down AI investment once they ponder what could happen to the economy if that investment wanes.


Wednesday, November 18, 2026

All Complex Systems Tend to Crash, Even With Redundancies and Backups

It might be normal to wonder whether there is something other than random failure going on when widely-disparate systems crash, such as Microsoft productivity apps; banking app functions,  air traffic control communication systems,  national security systems or other software. 


Eight outages in six days will do that.


Engineers of complex systems will not be surprised, as any complex system, especially when relying on software, is very hard to keep running “all the time.” Engineers refer to “availability” of components and systems as the percentage of time a component or full system is available for use. 


And the reliability of a system made up of many components can be substantially lower than the reliability of any individual component.


Reliability of each component

Probability entire system works

System availability

99%

Effectively zero

~0%

99.9%

Effectively zero

~0%

99.99%

0.0045%

~0%

99.999%

0.674%

0.67%

99.9999%

90.48%

90.48%

99.99999%

99.005%

99.0%

99.999999%

99.900%

99.90%


The implication: even if every individual component is 99.9999 percent reliable, the complete system is only about 90.5 percent reliable.


Of course, what really matters for whole-system reliability is whether component failures are independent, and whether the system can continue operating when individual components fail.


Assume a system has 100,000 potential points of failure, and that:

  • Each point has an equal probability of failure.

  • Failures are statistically independent.

  • Failure of any one point causes the entire system to fail.

  • All components must function for the system to operate.


For a system with 100,000 independent components and some amount of component dependency, the availability of each component has to be ery high to reach even 90-percent system availability, and super-high to reach system availability at a level when the whole system is unavailable only a few minutes a year.

Desired overall system reliability

Required reliability per component

90%

99.999895%

95%

99.999949%

99%

99.999990%

99.9%

99.999999%

99.99%

99.9999999%


To be sure, most complex systems are designed to avoid situations where there are single points of failure. So we build backup systems. We try to create architectures that protect against single component failures causing cascading outages. 


Consider an example closer to any typical user’s experience of using the internet and apps accessed from the internet. That is a complex system involving your own power supply, your devices, the apps you use and how often you use them. Your experience might also be affected by the way you create backups and redundancies. 


Still, a typical U.S. internet access user can expect the internet to be available and usable well over 99 percent of the time, but the reliability of any particular website, application, transaction, or online activity is lower.


What that means is that a typical user, using a system that is 99-percent available, will still experience more than seven hours of downtime every month. 


Availability

Downtime per year

Downtime per month

99%

3.65 days

7.3 hours

99.5%

1.83 days

3.65 hours

99.9%

8.76 hours

43.8 minutes

99.95%

4.38 hours

21.9 minutes

99.99%

52.6 minutes

4.4 minutes

99.999%

5.26 minutes

26 seconds



Element

Illustrative availability

Downtime/month

Downtime/year

Household electrical power

99.95%

22 min

4.4 hrs

Computer/device hardware

99.9%

44 min

8.8 hrs

Operating system/browser

99.9%

44 min

8.8 hrs

Home router/Wi-Fi

99.5%

3.65 hrs

43.8 hrs

Broadband ISP access

99.9%

44 min

8.8 hrs

ISP backbone/routing

99.95%

22 min

4.4 hrs

DNS and internet interconnection

99.95%

22 min

4.4 hrs

Cloud hosting infrastructure

99.95%

22 min

4.4 hrs

Individual applications

99.9%

44 min

8.8 hrs


The point is that any complex system is going to crash, sometimes, even when redundancies and backups are part of the architecture. And some use cases do not justify engineering much-higher levels of redundancy and backup. 


How many homes have two or more internet access provider accounts, just in case one supplier has an outage? How many routinely keep backup PCs, tablets and smartphones at the ready? How many have power generators? 


And most other parts of the experience are not controllable to any great extent. 


So it might seem as though multiple complex system failures are the result of some deliberate action and actor. Maybe not. We simply live in a world reliant on many complex systems, all of which have some propensity to crash.


Every AI Model Has to Make Value Choices

Arrow’s impossibility theorem suggests the impossibility of aligning human values with any single artificial intelligence model, for the si...